pool3d_fwd_common.hpp

// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#pragma once

#include <iostream>

#include "ck/ck.hpp"
#include "ck/utility/reduction_enums.hpp"
#include "ck/utility/reduction_functions_accumulate.hpp"
#include "ck/tensor_operation/gpu/device/reduction_operator_mapping.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_pool3d_fwd_ndhwc_ndhwc.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"

#include "ck/library/utility/check_err.hpp"
#include "ck/library/utility/device_memory.hpp"
#include "ck/library/utility/host_tensor.hpp"
#include "ck/library/utility/host_tensor_generator.hpp"
#include "ck/library/utility/literals.hpp"

template <typename InDataType,
          typename OutDataType,
          typename AccDataType,
          typename IndexDataType,
          ck::ReduceTensorOp ReduceOpId,
          bool PropagateNan,
          bool OutputIndex>
static void pool3d_host_verify(const Tensor<InDataType>& in,
                               Tensor<OutDataType>& out,
                               Tensor<IndexDataType>& out_indices,
                               const std::array<ck::index_t, 3>& window_spatial_lengths,
                               const std::array<ck::index_t, 3>& window_strides,
                               const std::array<ck::index_t, 3>& in_left_pads,
                               const std::array<ck::index_t, 3>& /*in_right_pads*/)
{
    const int32_t reduceLength =
        window_spatial_lengths[0] * window_spatial_lengths[1] * window_spatial_lengths[2];

    using ReduceOperation = typename ck::reduce_binary_operator<ReduceOpId>::opType;

    auto elementwise_ops =
        ck::reduce_unary_operator<ReduceOpId, true, true>::GetElementwiseOperator(reduceLength);

    auto in_elementwise_op  = std::get<0>(elementwise_ops);
    auto acc_elementwise_op = std::get<1>(elementwise_ops);

    if constexpr(!OutputIndex)
    {
        using Accumulation =
            ck::detail::AccumulateWithNanCheck<PropagateNan, ReduceOperation, AccDataType>;

        auto f_ncdhw = [&](auto n, auto c, auto do_, auto ho, auto wo) {
            auto accuVal = ReduceOperation::template GetIdentityValue<AccDataType>();

            for(ck::index_t z = 0; z < window_spatial_lengths[0]; ++z)
            {
                ck::index_t di = do_ * window_strides[0] + z - in_left_pads[0];
                for(ck::index_t y = 0; y < window_spatial_lengths[1]; ++y)
                {
                    ck::index_t hi = ho * window_strides[1] + y - in_left_pads[1];
                    for(ck::index_t x = 0; x < window_spatial_lengths[2]; ++x)
                    {
                        ck::index_t wi = wo * window_strides[2] + x - in_left_pads[2];
                        if(di >= 0 && di < static_cast<ck::index_t>(in.mDesc.GetLengths()[2]) &&
                           hi >= 0 && hi < static_cast<ck::index_t>(in.mDesc.GetLengths()[3]) &&
                           wi >= 0 && wi < static_cast<ck::index_t>(in.mDesc.GetLengths()[4]))
                        {
                            AccDataType currVal = static_cast<AccDataType>(in(n, c, di, hi, wi));

                            in_elementwise_op(currVal, currVal);

                            Accumulation::Calculate(accuVal, currVal);
                        }
                    }
                }
            }

            acc_elementwise_op(accuVal, accuVal);
            out(n, c, do_, ho, wo) = accuVal;
        };

        make_ParallelTensorFunctor(f_ncdhw,
                                   out.mDesc.GetLengths()[0],
                                   out.mDesc.GetLengths()[1],
                                   out.mDesc.GetLengths()[2],
                                   out.mDesc.GetLengths()[3],
                                   out.mDesc.GetLengths()[4])(std::thread::hardware_concurrency());
    }
    else
    {
        using Accumulation = ck::detail::AccumulateWithIndexAndNanCheck<PropagateNan,
                                                                        ReduceOperation,
                                                                        AccDataType,
                                                                        IndexDataType>;

        auto f_ncdhw = [&](auto n, auto c, auto do_, auto ho, auto wo) {
            auto accuVal            = ReduceOperation::template GetIdentityValue<AccDataType>();
            IndexDataType accuIndex = 0;

            for(ck::index_t z = 0; z < window_spatial_lengths[0]; ++z)
            {
                ck::index_t di = do_ * window_strides[0] + z - in_left_pads[0];
                for(ck::index_t y = 0; y < window_spatial_lengths[1]; ++y)
                {
                    ck::index_t hi = ho * window_strides[1] + y - in_left_pads[1];
                    for(ck::index_t x = 0; x < window_spatial_lengths[2]; ++x)
                    {
                        ck::index_t wi = wo * window_strides[2] + x - in_left_pads[2];
                        if(di >= 0 && di < static_cast<ck::index_t>(in.mDesc.GetLengths()[2]) &&
                           hi >= 0 && hi < static_cast<ck::index_t>(in.mDesc.GetLengths()[3]) &&
                           wi >= 0 && wi < static_cast<ck::index_t>(in.mDesc.GetLengths()[4]))
                        {
                            AccDataType currVal = static_cast<AccDataType>(in(n, c, di, hi, wi));
                            IndexDataType currIndex =
                                z * window_spatial_lengths[1] * window_spatial_lengths[2] +
                                y * window_spatial_lengths[2] + x;

                            in_elementwise_op(currVal, currVal);

                            Accumulation::Calculate(accuVal, currVal, accuIndex, currIndex);
                        }
                    }
                }
            }

            acc_elementwise_op(accuVal, accuVal);

            out(n, c, do_, ho, wo)         = accuVal;
            out_indices(n, c, do_, ho, wo) = accuIndex;
        };

        make_ParallelTensorFunctor(f_ncdhw,
                                   out.mDesc.GetLengths()[0],
                                   out.mDesc.GetLengths()[1],
                                   out.mDesc.GetLengths()[2],
                                   out.mDesc.GetLengths()[3],
                                   out.mDesc.GetLengths()[4])(std::thread::hardware_concurrency());
    };
}

template <typename InDataType,
          typename OutDataType,
          typename AccDataType,
          typename IndexDataType,
          typename InLayout,
          typename OutLayout,
          ck::ReduceTensorOp ReduceOpId,
          bool PropagateNan,
          bool OutputIndex>
bool pool3d_test(bool do_verification,
                 bool time_kernel,
                 ck::index_t N,
                 ck::index_t C,
                 ck::index_t Z,
                 ck::index_t Y,
                 ck::index_t X,
                 ck::index_t Di,
                 ck::index_t Hi,
                 ck::index_t Wi,
                 ck::index_t window_stride_d,
                 ck::index_t window_stride_h,
                 ck::index_t window_stride_w,
                 ck::index_t in_left_pad_d,
                 ck::index_t in_left_pad_h,
                 ck::index_t in_left_pad_w,
                 ck::index_t in_right_pad_d,
                 ck::index_t in_right_pad_h,
                 ck::index_t in_right_pad_w)
{
    using DevicePoolFwdInstance =
        ck::tensor_operation::device::DevicePool3dFwd_Input_N_Di_Hi_Wi_C_Output_N_Do_Ho_Wo_C<
            InDataType,  // InDataType
            OutDataType, // OutDataType
            AccDataType, // AccDataType
            ReduceOpId,
            OutputIndex,
            64, // BlockSize
            64, // ReduceMThreadClusterSize
            1,  // ReduceKThreadClusterSize
            4,  // ReduceMThreadSliceSize
            1,  // ReduceKThreadSliceSize
            4>; // InSrcOutDstVectorSize

    const ck::index_t Do = (Di + in_left_pad_d + in_right_pad_d - Z) / window_stride_d + 1;
    const ck::index_t Ho = (Hi + in_left_pad_h + in_right_pad_h - Y) / window_stride_h + 1;
    const ck::index_t Wo = (Wi + in_left_pad_w + in_right_pad_w - X) / window_stride_w + 1;

    const std::array<ck::index_t, 3> window_spatial_lengths{{Z, Y, X}};
    const std::array<ck::index_t, 3> window_strides{
        {window_stride_d, window_stride_h, window_stride_w}};
    const std::array<ck::index_t, 3> input_left_pads{{in_left_pad_d, in_left_pad_h, in_left_pad_w}};
    const std::array<ck::index_t, 3> input_right_pads{
        {in_right_pad_d, in_right_pad_h, in_right_pad_w}};

    // tensor layout
    auto f_host_tensor_descriptor = [](std::size_t N_,
                                       std::size_t C_,
                                       std::size_t D,
                                       std::size_t H,
                                       std::size_t W,
                                       auto layout) {
        using namespace ck::literals;

        if constexpr(ck::is_same<decltype(layout), ck::tensor_layout::convolution::NCDHW>::value)
        {
            return HostTensorDescriptor({N_, C_, D, H, W},
                                        {C_ * D * H * W, D * H * W, H * W, W, 1_uz});
        }
        else if constexpr(ck::is_same<decltype(layout),
                                      ck::tensor_layout::convolution::NDHWC>::value)
        {
            return HostTensorDescriptor({N_, C_, D, H, W},
                                        {D * C_ * H * W, 1_uz, C_ * H * W, W * C_, C_});
        }
    };

    Tensor<InDataType> in_n_c_di_hi_wi(f_host_tensor_descriptor(N, C, Di, Hi, Wi, InLayout{}));
    Tensor<OutDataType> out_n_c_do_ho_wo_host(
        f_host_tensor_descriptor(N, C, Do, Ho, Wo, OutLayout{}));
    Tensor<IndexDataType> out_indices_n_c_do_ho_wo_host(
        f_host_tensor_descriptor(N, C, Do, Ho, Wo, OutLayout{}));
    Tensor<OutDataType> out_n_c_do_ho_wo_device(
        f_host_tensor_descriptor(N, C, Do, Ho, Wo, OutLayout{}));
    Tensor<IndexDataType> out_indices_n_c_do_ho_wo_device(
        f_host_tensor_descriptor(N, C, Do, Ho, Wo, OutLayout{}));

    std::cout << "in_n_c_di_hi_wi: " << in_n_c_di_hi_wi.mDesc << std::endl;
    std::cout << "out_n_c_do_ho_wo: " << out_n_c_do_ho_wo_host.mDesc << std::endl;

    in_n_c_di_hi_wi.GenerateTensorValue(GeneratorTensor_3<InDataType>{-1.0, 1.0});

    DeviceMem in_device_buf(sizeof(InDataType) * in_n_c_di_hi_wi.mDesc.GetElementSpaceSize());
    DeviceMem out_device_buf(sizeof(OutDataType) *
                             out_n_c_do_ho_wo_device.mDesc.GetElementSpaceSize());
    DeviceMem out_indices_device_buf(sizeof(IndexDataType) *
                                     out_indices_n_c_do_ho_wo_device.mDesc.GetElementSpaceSize());

    in_device_buf.ToDevice(in_n_c_di_hi_wi.mData.data());

    auto pool         = DevicePoolFwdInstance{};
    auto invoker_ptr  = pool.MakeInvokerPointer();
    auto argument_ptr = pool.MakeArgumentPointer(
        static_cast<InDataType*>(in_device_buf.GetDeviceBuffer()),
        static_cast<OutDataType*>(out_device_buf.GetDeviceBuffer()),
        static_cast<IndexDataType*>(out_indices_device_buf.GetDeviceBuffer()),
        {N, C, Di, Hi, Wi},
        {Z, Y, X},
        {N, C, Do, Ho, Wo},
        window_strides,
        input_left_pads,
        input_right_pads,
        {2, 3, 4});

    if(!pool.IsSupportedArgument(argument_ptr.get()))
    {
        throw std::runtime_error("wrong! device_op with the specified compilation parameters does "
                                 "not support this problem");
    }

    float ave_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
    std::cout << "Perf: " << ave_time << std::endl;

    bool pass = true;

    if(do_verification)
    {
        pool3d_host_verify<InDataType,
                           OutDataType,
                           AccDataType,
                           IndexDataType,
                           ReduceOpId,
                           PropagateNan,
                           OutputIndex>(in_n_c_di_hi_wi,
                                        out_n_c_do_ho_wo_host,
                                        out_indices_n_c_do_ho_wo_host,
                                        window_spatial_lengths,
                                        window_strides,
                                        input_left_pads,
                                        input_right_pads);

        out_device_buf.FromDevice(out_n_c_do_ho_wo_device.mData.data());

        pass = pass && ck::utils::check_err(out_n_c_do_ho_wo_device, out_n_c_do_ho_wo_host);

        if constexpr(OutputIndex)
        {
            out_indices_device_buf.FromDevice(out_indices_n_c_do_ho_wo_device.mData.data());

            pass = pass && ck::utils::check_err(out_indices_n_c_do_ho_wo_device,
                                                out_indices_n_c_do_ho_wo_host);
        };
    }

    return (pass);
};